According to @aloctavodia, we can update $n$, where $n < m$, trees during sampling since there is not much statistical difference in the results. In the code, we control $n$ with batch_tune and batch_post—both default to 10% of the trees. During tuning we would store all $m$ trees but during sampling use $m \cdot \text{batch} = B$ trees
Below is an illustrative figure of the approach that displays the sliding window, again with 10% of the trees:
According to @aloctavodia, we can update$n$ , where $n < m$ , trees during sampling since there is not much statistical difference in the results. In the code, we control $n$ with $m$ trees but during sampling use $m \cdot \text{batch} = B$ trees
batch_tuneandbatch_post—both default to 10% of the trees. During tuning we would store allBelow is an illustrative figure of the approach that displays the sliding window, again with 10% of the trees:
